AI for Safe Climate Cooling: Deep Learning and XAI for Rapid SRM Risk Assessment

As global warming accelerates, Solar Radiation Management (SRM), including stratospheric aerosol injection (SAI) and marine cloud brightening (MCB), are increasingly viewed as a potential tool to circumvent near-term climate risks. However, the complex interplay between chemistry, radiation, and dynamics creates deep uncertainties regarding regional climate disruptions and unintended feedback. Traditional evaluation relies on computationally intensive Earth System Models (ESMs), which often limit the scalability, spatial resolutions and temporal scales required for adaptive governance.

To bridge this gap, we propose a framework that leverages deep learning and explainable artificial intelligence (XAI) to accelerate the assessment of SRM impacts. A core component of our work involves adapting NeuralGCM, a hybrid atmospheric model that combines differentiable physics with machine learning, to SRM-specific scenarios. By training on diverse climate model simulations and historical analogs, we aim to assess lower temporal and spatial resolution, providing high-fidelity results of traditional models at a fraction of the computational cost.

To ensure these "black-box" models are reliable for policy-relevant science, we adapt XAI methods specifically for the climate context. These tools allow domain experts to interpret model behavior across multiple timescales, detect incorrectly learned physical mechanisms or spurious correlations, and assess risk propagation and regional uncertainties, particularly in vulnerable areas. By improving the speed, transparency, and reliability of climate intervention modeling, this approach contributes to a safer, more informed exploration of SRM as a component of global climate strategy.